ROIAlign

Versioned name: ROIAlign-9

Category: Object detection

Short description: ROIAlign is a pooling layer used over feature maps of non-uniform input sizes and outputs a feature map of a fixed size.

Detailed description: Reference.

ROIAlign performs the following for each Region of Interest (ROI) for each input feature map:

1. Multiply box coordinates with spatial_scale to produce box coordinates relative to the input feature map size based on aligned_mode attribute.

2. Divide the box into bins according to the sampling_ratio attribute.

3. Apply bilinear interpolation with 4 points in each bin and apply maximum or average pooling based on mode attribute to produce output feature map element.

Attributes

• pooled_h

  • Description: pooled_h is the height of the ROI output feature map.

  • Range of values: a positive integer

  • Type: int

  • Required: yes

• pooled_w

  • Description: pooled_w is the width of the ROI output feature map.

  • Range of values: a positive integer

  • Type: int

  • Required: yes

• sampling_ratio

  • Description: sampling_ratio is the number of bins over height and width to use to calculate each output feature map element. If the value is equal to 0 then use adaptive number of elements over height and width: ceil(roi_height / pooled_h) and ceil(roi_width / pooled_w) respectively.

  • Range of values: a non-negative integer

  • Type: int

  • Required: yes

• spatial_scale

  • Description: spatial_scale is a multiplicative spatial scale factor to translate ROI coordinates from their input spatial scale to the scale used when pooling.

  • Range of values: a positive floating-point number

  • Type: float

  • Required: yes

• mode

  • Description: mode specifies a method to perform pooling to produce output feature map elements.

  • Range of values:

    • max - maximum pooling

    • avg - average pooling

  • Type: string

  • Required: yes

• aligned_mode

  • Description: aligned_mode specifies how to transform the coordinate in original tensor to the resized tensor.

  • Range of values: name of the transformation mode in string format (here spatial_scale is resized_shape[x] / original_shape[x], resized_shape[x] is the shape of resized tensor in axis x, original_shape[x] is the shape of original tensor in axis x and x_original is a coordinate in axis x, for any axis x from the input axes):

    • asymmetric - the coordinate in the resized tensor axis x is calculated according to the formula x_original * spatial_scale

    • half_pixel_for_nn - the coordinate in the resized tensor axis x is x_original * spatial_scale - 0.5

    • half_pixel - the coordinate in the resized tensor axis x is calculated as ((x_original + 0.5) * spatial_scale) - 0.5

  • Type: string

  • Default value: asymmetric

  • Required: no

Inputs:

• 1: 4D input tensor of shape [N, C, H, W] with feature maps of type T. Required.

• 2: 2D input tensor of shape [NUM_ROIS, 4] describing box consisting of 4 element tuples: [x_1, y_1, x_2, y_2] in relative coordinates of type T. The box height and width are calculated the following way:

  • If aligned_mode equals asymmetric: roi_width = max(spatial_scale * (x_2 - x_1), 1.0), roi_height = max(spatial_scale * (y_2 - y_1), 1.0), so the malformed boxes are expressed as a box of size 1 x 1.

  • else: roi_width = spatial_scale * (x_2 - x_1), roi_height = spatial_scale * (y_2 - y_1).

  • Required.

• 3: 1D input tensor of shape [NUM_ROIS] with batch indices of type IND_T. Required.

Outputs:

• 1: 4D output tensor of shape [NUM_ROIS, C, pooled_h, pooled_w] with feature maps of type T.

Types

• T: any supported floating-point type.

• IND_T: any supported integer type.

Example

<layer ... type="ROIAlign" ... >
    <data pooled_h="6" pooled_w="6" spatial_scale="16.0" sampling_ratio="2" mode="avg" aligned_mode="half_pixel"/>
    <input>
        <port id="0">
            <dim>7</dim>
            <dim>256</dim>
            <dim>200</dim>
            <dim>200</dim>
        </port>
        <port id="1">
            <dim>1000</dim>
            <dim>4</dim>
        </port>
        <port id="2">
            <dim>1000</dim>
        </port>
    </input>
    <output>
        <port id="3" precision="FP32">
            <dim>1000</dim>
            <dim>256</dim>
            <dim>6</dim>
            <dim>6</dim>
        </port>
    </output>
</layer>